🔍 Read the full analysis: Meta Muse Ember: 36 Points For $0.14 A Task, With Frequent Refusals on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Meta released Muse Ember on November 9, and Artificial Analysis scored its max-effort version 36 on its Intelligence Index at an estimated $0.14 per index task. The same source reports frequent refusals on a factual-knowledge test, alongside a lower hallucination rate than several compared models; the results are benchmark measurements, not guarantees for other workloads.
Meta released Muse Ember on November 9, and independent model evaluator Artificial Analysis scored its max-effort version 36 on the organization’s Intelligence Index, at an estimated $0.14 per index task. The evaluator reports that Claude Haiku 5.5 also scores 36 but costs $0.41 per task in the comparison, while Muse Ember often declines to answer questions on a factual-knowledge test.
Artificial Analysis says Muse Ember is the efficient tier of Meta’s Muse model line, available through Meta’s API and the Muse Code coding agent. It accepts text and images and generates text. The evaluator tested two reasoning-effort settings: the max version scored 36 for $0.14 per index task, while xhigh scored 34 for $0.12. These task costs represent average API costs for tasks in the index, not a universal price for every customer request.
The source lists prices of $0.10 per million input tokens and $0.50 per million output tokens, matching its listed rates for Haiku 5.5 and GPT-6 Luna. Cached input is priced at $0.004 per million tokens, which Artificial Analysis describes as 60% below Haiku’s rate. The evaluator attributes Ember’s lower index-task cost, despite matching Haiku’s listed token prices, to cheaper cached input and its use of roughly a third as many output tokens as Haiku across the index.
Results vary by test. Artificial Analysis reports Muse Ember at 68.1% on AutomationBench-AA, level with Meta’s higher-priced Muse Wonder and above Haiku 5.5’s 35.4%. Ember scored 40.9% on Terminal-Bench 4.0, compared with Haiku’s 32.8%, but trailed it on FrontierCode 2.0, scoring 27.2% to Haiku’s 34.9%. In the evaluator’s coding-agent index, Ember in Muse Code scored 46, compared with 37 for Haiku in Claude Code and 31 for GPT-6 Luna in Codex.
Meta Muse Ember: 36 Points for $0.14 a Task
Artificial Analysis reports a low benchmark task cost and frequent refusals on a factual-knowledge test. These measurements compare models under specific test conditions, not every real-world workload.
What the comparison shows
Ember and Claude Haiku 5.5 both scored 36, but benchmark costs and results differed across tests.
| Model or test | Ember | Comparison |
|---|---|---|
| Index task cost | $0.14 | Haiku 5.5: $0.41 |
| AutomationBench-AA | 68.1% | Haiku 5.5: 35.4% |
| FrontierCode 2.0 | 27.2% | Haiku 5.5: 34.9% |
Strengths and trade-offs
- Lower reported index-task cost than Haiku 5.5.
- Strong result on AutomationBench-AA in this comparison.
- Ember in Muse Code scored 46 on the coding-agent index.
- Frequent refusals may interrupt workflows requiring direct answers.
- Factual recall was weak on the tested knowledge benchmark.
- Ember trailed Haiku 5.5 on FrontierCode 2.0.
Before using it at work
Benchmark costs and scores do not predict results on a team’s own prompts, data, or usage.
Compare Ember with alternatives using representative workloads.
Try retrieved documents or search results instead of relying on model memory.
Track accuracy and abstentions; consider review or a fallback model.
Lower Costs Come With Trade-Offs
The reported price gap could matter to organizations running high-volume, repetitive tasks, where small differences in per-task costs can affect overall API spending. Artificial Analysis estimates that the index tasks cost about one-third as much with Muse Ember as with Haiku 5.5. That comparison applies to the evaluator’s benchmark setup; it does not establish the savings a business would achieve on its own prompts, data, or usage pattern.
The reported refusals also shape where the model may fit. On Artificial Analysis’s AA-Omniscience test, Ember answered 21.1% of questions correctly and declined 64%; its reported hallucination rate was 17.4%. Haiku 5.5 answered 34.7% correctly and had a 34.2% hallucination rate in the same comparison. These results suggest a model that may be more likely to abstain than to invent an answer on that test, but frequent refusals can also interrupt workflows that require a direct response.
For readers choosing models, the figures point to a practical distinction: Ember’s benchmark strengths include workflow automation and some coding measures, while its tested factual recall is weak. The source recommends evaluating the model on a team’s own tasks and providing relevant information—such as documents or retrieved search results—rather than relying on model memory. That is an assessment based on the cited results, not a guarantee of performance in production.
Products Worth Considering
Apple Mac mini (2026), M6 12-core CPU and GPU, 24GB Memory, 1TB SSD
Logitech MX Master 4 for Mac Wireless Ergonomic Mouse, White Silver
FlexiSpot E6 Electric Standing Desk, 55 × 28 in, Bamboo Top, White Frame
As an affiliate, we earn on qualifying purchases.
How the Benchmark Compares Models
Artificial Analysis’s Intelligence Index combines nine evaluations, according to the supplied source. Its cost-per-task figure is the average API cost for one task in that index. The index score and cost are therefore useful for comparing models under the evaluator’s test conditions, but they do not measure every capability or predict the price of an individual application.
The source compares Muse Ember (max) with several models at different price and score points. Haiku 5.5 (max) is listed at the same score of 36 and $0.41 per task. Meta’s Muse Wonder scores 44 at $6.95 per task. Anthropic’s Claude Fable 5.2 at its highest effort setting scores 65 at $26.54 per task. Those figures show the source’s reported cost-and-score trade-offs; they do not establish that models with different scores can substitute for one another on a particular job.
Artificial Analysis also reports 440 output tokens per second and 3.2 minutes per index task for Ember at max effort, with a one-million-token context window. Its xhigh version is listed at 423 output tokens per second and 2.8 minutes per task. These are evaluator-reported measurements, and actual speed can depend on the task and service conditions.
“Knows less, guesses less.”
— Thorsten Meyer AI, describing the benchmark results
Limits of the Reported Results
The figures in the source are attributed to Artificial Analysis model pages and tests accessed November 9 and 10, 2026. The supplied material does not include the full methodology, confidence ranges, or independent replication details for each benchmark result. It is also not clear how representative the index tasks are of customer workloads or whether API costs will remain at the listed levels.
The AA-Omniscience figures describe performance on one factual-knowledge test. They do not establish how often Ember will refuse in a particular application, or whether its reported hallucination rate will hold when prompts, tools, and supplied context differ. The source also does not provide enough detail to assess the quality of answers when Ember does respond across all use cases.
Test Results on Real Workloads
The next practical step for prospective users is to compare Muse Ember with alternatives on their own tasks and cost assumptions. For applications that rely on external information, teams can test it with retrieved documents or database results and track both answer accuracy and refusal rates. Workflows that need a response every time may also need a review path or a fallback model when Ember declines.
Meta’s API and Muse Code are the stated access routes in the source material. It does not give a schedule for further releases or updates, so no next product milestone is confirmed. Users will need to watch for additional benchmark testing, product documentation and pricing updates before drawing broader conclusions about the model’s performance or operating costs.
Key Questions
What score did Meta Muse Ember receive?
Artificial Analysis reports a score of 36 for Muse Ember at max reasoning effort on its Intelligence Index. Its xhigh version scored 34 in the source’s comparison.
How much does Muse Ember cost per index task?
The source lists an average of $0.14 per index task for the max version and $0.12 for xhigh. These are benchmark task-cost estimates, not a flat price for every request.
Why are refusals part of the story?
On Artificial Analysis’s AA-Omniscience test, Muse Ember reportedly declined 64% of questions and answered 21.1% correctly. The same test reported a 17.4% hallucination rate, but results may differ on other tasks.
Is Muse Ember better than Claude Haiku 5.5?
The source does not support a single overall winner. Both scored 36 on the Intelligence Index in the comparison, while Ember had a lower reported index-task cost. Their results differed across workflow, coding and factual-knowledge tests.
Where can users access Muse Ember?
The source says Meta made Muse Ember available through Meta’s API and Muse Code, Meta’s coding agent. It does not provide further access or availability details.
Source: ThorstenMeyerAI.com
